Enhanced CNN-Bilstm Transfer Learning Based Method on Medical Image Diagnosis

Abstract

This study demonstrates the strength of deep leaning models to classify chest X-rays images of COVID-19 and Tuberculosis datasets using transfer learning. Convolutional Neural Network (CNN) of ImageNet is adopted as pretrained model with data augmentation and contrast medical images enhancement. The images are learned and trained using transfer learning for feature representations to identify categories of 10,890 chest X-ray images with hyperparameters to prevent overfitting. Three pretrained CNN models; VGG16, ResNet50 and Inceptionv3 are employed for evaluation of the model, the output of the CNNs pretrained models are fed into Bidirectional Long Short Term Memory (BiLSTM). This performs sequence problems and contributes to the previous task on transfer learning to achieve better classification. The best performance is achieved using a combination of features extracted from CNN-based models and BiLSTM with increase in the accuracy, precision, recall, and F1 score for training datasets from epoch 50 to 250 with values 0.62 to 0.99, 0.60 to 0.84, 0.54 to 0.88 and 0.52 to 0.85 respectively. Inceptionv3 shows the best performance while ResNet50 with low performance. It shows that there is improved in the image enhancement with augmentation for accuracy with considerable value of 0.002 for loss compared to existing CNNs pre-trained.

Keywords— BiLSTM; Chest x-ray; Classification; CNN pretrained; Transfer learning